memU and OpenClaw: persistent agent memory explained

This Italian-language explainer examines how an AI agent keeps useful information across long conversations and fresh sessions. The speaker starts with context window limits and context rot, then explains compaction: summarizing a conversation frees space but can discard details. Codex and Claude are examples in this discussion.

The central comparison is between RAG with a vector database and agentic file search. In a PDF example, the agent reads an index, selects page 10 and requests only the relevant text through tools. The speaker argues that this can avoid some retrieval misses, while acknowledging that semantic search still has uses and can complement direct file access.

OpenClaw provides the most detailed memory example. The presenter describes dated Markdown logs for daily activity and a separate memory file for durable preferences and decisions. New sessions read recent logs alongside that persistent file. Before compaction, a memory flush prompts the model to save lasting notes to disk. Semantic search helps locate older files, which the agent then reads directly.

The memU section describes memory as a hierarchical file system: topics become folders, remembered items become files, and references connect them. The speaker walks through preference, relationship, interaction-history, knowledge and recent-context files that an agent can read through tools. This is an architecture explanation, not a memU installation or measured long-running test. Ralph Loop illustrates how requirements and progress files carry work between coding sessions. Persistent local storage does not establish that the underlying models run offline. The speaker closes with his Rizzo Clow teaching project, where memory improvements and file permission sandboxing remain in progress.